Morning Edition · Wednesday, July 22, 2026Published at 1:47 AM EDT · New York

Z.ai Powers Up a 1-Gigawatt AI Training Cluster With No Nvidia Chips Inside

The GLM developer now runs multiple clusters of more than 10,000 domestic accelerators each, a year after Washington blacklisted it.

Save

Z.ai Powers Up a 1-Gigawatt AI Training Cluster With No Nvidia Chips Inside

Z.ai, the Chinese lab formerly known as Zhipu, has finished building a data center that draws roughly one gigawatt of power and is stocked entirely with domestically made accelerators. It has already begun operating part of the site, Bloomberg reported citing a person familiar with the project. Russian-language technology channels carried the same account, noting that the site will train the company's GLM model family and that Z.ai now operates several compute clusters holding more than 10,000 chips each.

The source did not name the chip supplier, but Z.ai's recent training history suggests Huawei. The company released GLM-5.2 in June, an open-weight model it says was trained without any Nvidia hardware, most likely on Huawei Ascend accelerators. Z.ai has had commercial reason to build a non-Nvidia stack since Washington added it to an export blacklist in January 2025.

The claims come from a single sourced report rather than an independent audit. The specifics that matter most for capability, meaning yield, interconnect bandwidth, and the effective utilization of a cluster this size on Chinese chips, are not disclosed. A gigawatt of nameplate power is not the same as a gigawatt of useful FLOPs. What is verified is the direction: a sanctioned Chinese lab is building frontier-scale compute without American chips and training competitive open-weight models on it.

Veracity: Plausible

72/100

If true, who benefits: Huawei's accelerator business and Beijing's tech-sovereignty narrative gain, and China-compute bulls betting the export-control thesis is failing, while Nvidia loses a market it was already barred from.
The nuance: The account traces to a single anonymous Bloomberg source with no independent audit, one gigawatt is nameplate power not delivered compute, the site is only partly operational, and the chip vendor, yield, and interconnect are undisclosed.

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

Export controls were designed to deny China the compute to train frontier models. If a blacklisted lab can assemble gigawatt-scale clusters on Huawei-class chips and ship open-weight GLM models from them, the control regime shifts from denying China compute to raising its cost, meaning Chinese labs pay more power and accept worse interconnect per unit of capability rather than being stopped. Nvidia loses a market it was already barred from, but the larger exposure is the policy premise that gating access slows the Chinese frontier.

What to watch

Whether GLM models trained on this cluster hold their benchmark position against US frontier models over the next few releases, which would show that domestic chips are production-viable rather than a one-time demonstration.
Any disclosure of the chip vendor and per-cluster interconnect, the detail that separates a real training fabric from a warehouse of idle accelerators.

Observations to monitor, not financial advice.

3 sources

Synthesized from: Polylog editors · Tom's Hardware · The Next Web

Part of a tracked trend

AI Sovereignty and Export Controls on Frontier Models

Over the next 3-6 months, governments increasingly treat frontier AI models as strategic national assets — extending export controls to model access itself and backing domestic 'champion' labs as sovereignty plays.